Rust
Summary - SIMD
Combine Simd<T, N> and Mask to build a vector sum.
SIMD summary
SIMD processes several lanes with one instruction. std::simd (Simd<T, N>, masks, reduce_*) is portable but nightly; the stable options are std::arch and auto-vectorization. Vertical operations are lane-wise, horizontal ones reduce the lanes. Mask + select give branchless conditions. For auto-vectorization: access without bounds checks, multiple accumulators, a branchless body; for wider vectors, target-cpu or #[target_feature].
Dot product with four accumulators
Vertical multiply-add, a single horizontal reduction at the end, and the leftovers handled separately:
fn nokta4(a: &[f32], b: &[f32]) -> f32 {
assert_eq!(a.len(), b.len());
let mut birikim = [0.0f32; 4];
for (x, y) in a.chunks_exact(4).zip(b.chunks_exact(4)) {
for i in 0..4 { birikim[i] += x[i] * y[i]; }
}
let n = a.len() / 4 * 4;
let kalan: f32 = a[n..].iter().zip(&b[n..]).map(|(x, y)| x * y).sum();
birikim.iter().sum::<f32>() + kalan
}
Exercise Fast dot product
Write the nokta4 function above. Let a be the numbers 1 to 10 (f32) and b ten 1.0s; print the result.
Show example solution
fn nokta4(a: &[f32], b: &[f32]) -> f32 {
assert_eq!(a.len(), b.len());
let mut birikim = [0.0f32; 4];
for (x, y) in a.chunks_exact(4).zip(b.chunks_exact(4)) {
for i in 0..4 {
birikim[i] += x[i] * y[i];
}
}
let n = a.len() / 4 * 4;
let kalan: f32 = a[n..].iter().zip(&b[n..]).map(|(x, y)| x * y).sum();
birikim.iter().sum::<f32>() + kalan
}
fn main() {
let a: Vec<f32> = (1..=10).map(|x| x as f32).collect();
let b = vec![1.0f32; 10];
println!("{}", nokta4(&a, &b));
}
Test yourself
-
Why are 4 separate accumulators fast in a dot product?
- It uses less memory
- It's more accurate
- It isn't fast
- Independent chains progress at once and map directly onto a 4-lane vector
Show answer
Independent chains progress at once and map directly onto a 4-lane vector — With a single accumulator each addition waits for the previous one; four accumulators hide the latency.
-
Pair the pieces
for (x, y) in a.chunks_exact(4).____(b.chunks_exact(4)) {
for i in 0..4 { birikim[i] += x[i] * y[i]; }
}
Show answer
for (x, y) in a.chunks_exact(4).zip(b.chunks_exact(4)) {
for i in 0..4 { birikim[i] += x[i] * y[i]; }
}
-
Group summary
- Simd<T, N>
- N-lane vector
- Mask + select
- Branchless condition
- reduce_sum
- Horizontal sum
- Multiple accumulators
- Vectorization friendly
Next lesson: asm! macro
